OpenAI keeps a public log of times its own models misbehaved: the Misalignment Reports and Notices page at https://alignment.openai.com/misalignment-reports/
As of October 26, 2026 the page lists 20 reports (it shows "20 of 20 reports") plus 3 separate notices. Reports were added in batches: 6 at launch on Sep 16, then 3 on Sep 25, 3 on Oct 2, 3 on Oct 9, 4 on Oct 15 and 1 on Oct 20. The latest one describes two internal coding agents weakening their own misalignment monitoring, one of which then denied doing it.
Question: how many reports will that page list at 12:00 PM Pacific on November 25, 2026?
Resolution:
I count the rows in the Reports table, using the page's own "N of N reports" counter with no filters applied. Notices are not counted.
A report counts once no matter how many incidents it covers. Updates to an existing report do not add to the count.
If a report is removed before the snapshot, it isn't counted.
If the page is down or restructured at that moment, I use the closest Wayback Machine (web.archive.org) snapshot taken before the deadline. If OpenAI moves the list to another official OpenAI URL, I count it there.
Buckets are inclusive.
I'll resolve shortly after the snapshot time.
My starting view: 20–22 about 15%, 23–25 about 22%, 26–28 about 25%, 29–31 about 20%, 32 or more about 18%.
Why so spread out: after the 6-report launch, OpenAI added 3, 3, 3 and 4 reports in four straight weekly batches, then only 1 on Oct 20, and none since. Most of those reports describe incidents from May and June, so they've been clearing a backlog. If the backlog is nearly empty, the count could stall in the low 20s. If they keep the weekly rhythm, it lands around 30.
Things that would move me: a new batch this week (then I'd shift up), or another quiet week (then I'd shift down). The page: https://alignment.openai.com/misalignment-reports/